adaptmach

AdaptMach is an autonomous machine-design agent. You describe a machine, like “deliver a pizza from San Francisco to Fremont through a hurricane”. The agent, running on Meta’s Muse Spark (muse-spark-1.3), writes a Python design() function that picks rotors, motors, battery chemistry, frame, ducts and cruise speed. Every design is graded by our own physics engine (physics.py), not a game engine. It uses momentum theory for rotor thrust and power, with real-rotor losses and duct augmentation, and real battery energy and C-rate limits. Mass builds up from components and feeds back into power needs. For missions it models forward-flight drag, headwind, gust stability, rain, and trip energy with reserve. Mars and Titan change gravity and air density. A Three.js 3D view flies each design through the storm exactly as the physics says: out of battery over the bay, overheating, flipped by a gust, or delivered. Failures feed back into the next attempt. After each job, the agent writes down what it learned, as a lesson or a Python tool. That lesson is tested on machines it has never seen, such as a jetpack on Mars or a supply drone in a storm. It’s kept only if those machines need fewer attempts; otherwise it’s rolled back. You can also prototype any machine live and watch each iteration as a 3D model. Stack: Python standard library only (simulator, sandboxed code execution, learning loop), Three.js front end, Meta Model API.